hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-27 11:24:05 +00:00
Code Issues Actions Projects Releases Wiki Activity

Updated bipred_recon to work faster

Browse source
This commit is contained in:
Reima Hyvönen 2018-08-28 15:12:31 +03:00
parent 47b357cca2
commit 7de5c74434
2 changed files with 27 additions and 83 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

94
src/strategies/avx2/picture-avx2.c
View file

@ -748,8 +748,8 @@ static void inter_recon_bipred_no_mov_avx2(
case 4:
sample0_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
sample0_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
temp_y_epi8 = _mm256_avg_epu8(sample0_epi8, sample1_epi8);
@ -807,33 +807,15 @@ static void inter_recon_bipred_no_mov_avx2(
switch (width)
{
case 4:
sample0_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&temp_lcu_u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&lcu->rec.u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
temp_u_epi8 = _mm256_avg_epu8(sample0_epi8, sample1_epi8);
sample0_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&temp_lcu_v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&lcu->rec.v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
temp_v_epi8 = _mm256_avg_epu8(sample0_epi8, sample1_epi8);
lcu->rec.u[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu + 0] = _mm256_extract_epi8(temp_u_epi8, 0);
lcu->rec.u[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu + 1] = _mm256_extract_epi8(temp_u_epi8, 1);
lcu->rec.v[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu + 0] = _mm256_extract_epi8(temp_v_epi8, 0);
lcu->rec.v[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu + 1] = _mm256_extract_epi8(temp_v_epi8, 1);
break;
case 8:
sample0_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&temp_lcu_u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&lcu->rec.u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample0_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&temp_lcu_u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&lcu->rec.u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
temp_u_epi8 = _mm256_avg_epu8(sample0_epi8, sample1_epi8);
sample0_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&temp_lcu_v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&lcu->rec.v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample0_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&temp_lcu_v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&lcu->rec.v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
temp_v_epi8 = _mm256_avg_epu8(sample0_epi8, sample1_epi8);
pointer = (int32_t*)&(lcu->rec.u[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu]);
@ -921,9 +903,8 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
kvz_pixel* temp_lcu_u,
kvz_pixel* temp_lcu_v)
{
bool test = false;
if (test)//(hi_prec_luma_rec0 == 0 && hi_prec_luma_rec1 == 0 && hi_prec_chroma_rec0 == 0 && hi_prec_chroma_rec1 == 0)
if (hi_prec_luma_rec0 == 0 && hi_prec_luma_rec1 == 0 && hi_prec_chroma_rec0 == 0 && hi_prec_chroma_rec1 == 0)
{
inter_recon_bipred_no_mov_avx2(height, width, ypos, xpos, high_precision_rec0, high_precision_rec1, lcu, temp_lcu_y, temp_lcu_u, temp_lcu_v);
}
@ -934,7 +915,7 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
int y_in_lcu, x_in_lcu;
int shift = 15 - KVZ_BIT_DEPTH;
int offset = 1 << (shift - 1);
int shift_left = 14 - KVZ_BIT_DEPTH;
int8_t shift_left = 14 - KVZ_BIT_DEPTH;
__m256i temp_epi16_y, temp_epi8, temp_y_epi16, sample0_epi16, sample1_epi16;
__m256i temp_u_epi16 = _mm256_setzero_si256();
__m256i temp_v_epi16 = _mm256_setzero_si256();
@ -965,12 +946,13 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
case 4:
// Load total of 8 elements from memory to vector (4 are used)
// Load total of 4 elements from memory to vector
sample0_epi16 = hi_prec_luma_rec0 ? _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*) &(high_precision_rec0->y[y_in_lcu * LCU_WIDTH + x_in_lcu]))) :
_mm256_slli_epi16(_mm256_cvtepu8_epi16((_mm_loadl_epi64((__m128i*) &(temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu])))), shift_left);
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_cvtsi32_si128(*(int32_t*)&(temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu]))), shift_left);
sample1_epi16 = hi_prec_luma_rec1 ? _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*) &(high_precision_rec1->y[y_in_lcu * LCU_WIDTH + x_in_lcu]))) :
_mm256_slli_epi16(_mm256_cvtepu8_epi16((_mm_loadl_epi64((__m128i*) &(lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu])))), shift_left);
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_cvtsi32_si128(*(int32_t*) &(lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu]))), shift_left);
// (sample1 + sample2 + offset)>>shift
temp_y_epi16 = _mm256_add_epi16(sample0_epi16, sample1_epi16);
@ -982,7 +964,6 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
pointer = (int32_t*)&(lcu->rec.y[(y_in_lcu)* LCU_WIDTH + x_in_lcu]);
*pointer = _mm_cvtsi128_si32(_mm256_castsi256_si128(temp_epi8));
//lcu->rec.y[(y_in_lcu)* LCU_WIDTH + x_in_lcu] = _mm256_extract_epi32(temp_epi8, 0);
break;
@ -1000,7 +981,7 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
temp_y_epi16 = _mm256_add_epi16(temp_y_epi16, offset_epi16);
temp_y_epi16 = _mm256_srai_epi16(temp_y_epi16, shift);
// Pack the bits from 1-bit to 8-bit
// Pack the bits from 16-bit to 8-bit
temp_epi8 = _mm256_packus_epi16(temp_y_epi16, temp_y_epi16);
@ -1067,52 +1048,15 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
x_in_lcu = (((xpos >> 1) + temp_x) & (LCU_WIDTH_C - 1));
switch (width) {
case 4:
sample0_epi16 = hi_prec_chroma_rec0 ? _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*) &(high_precision_rec0->u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))) :
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_loadl_epi64((__m128i*) &(temp_lcu_u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
sample1_epi16 = hi_prec_chroma_rec1 ? _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*) &(high_precision_rec1->u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))) :
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_loadl_epi64((__m128i*) &(lcu->rec.u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
// (sample1 + sample2 + offset)>>shift
temp_u_epi16 = _mm256_add_epi16(sample0_epi16, sample1_epi16);
temp_u_epi16 = _mm256_add_epi16(temp_u_epi16, offset_epi16);
temp_u_epi16 = _mm256_srai_epi16(temp_u_epi16, shift);
sample0_epi16 = hi_prec_chroma_rec0 ? _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*) &(high_precision_rec0->v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))) :
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_loadl_epi64((__m128i*) &(temp_lcu_v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
sample1_epi16 = hi_prec_chroma_rec1 ? _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*) &(high_precision_rec1->v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))) :
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_loadl_epi64((__m128i*) &(lcu->rec.v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
// (sample1 + sample2 + offset)>>shift
temp_v_epi16 = _mm256_add_epi16(sample0_epi16, sample1_epi16);
temp_v_epi16 = _mm256_add_epi16(temp_v_epi16, offset_epi16);
temp_v_epi16 = _mm256_srai_epi16(temp_v_epi16, shift);
temp_epi8 = _mm256_packus_epi16(temp_u_epi16, temp_u_epi16);
lcu->rec.u[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu + 0] = _mm256_extract_epi8(temp_epi8, 0);
lcu->rec.u[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu + 1] = _mm256_extract_epi8(temp_epi8, 1);
temp_epi8 = _mm256_packus_epi16(temp_v_epi16, temp_v_epi16);
lcu->rec.v[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu + 0] = _mm256_extract_epi8(temp_epi8, 0);
lcu->rec.v[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu + 1] = _mm256_extract_epi8(temp_epi8, 1);
break;
switch (width) {
case 8:
sample0_epi16 = hi_prec_chroma_rec0 ? _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*) &(high_precision_rec0->u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))) :
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_loadl_epi64((__m128i*) &(temp_lcu_u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_cvtsi32_si128(*(int32_t*) &(temp_lcu_u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
sample1_epi16 = hi_prec_chroma_rec1 ? _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*) &(high_precision_rec1->u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))) :
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_loadl_epi64((__m128i*) &(lcu->rec.u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_cvtsi32_si128(*(int32_t*) &(lcu->rec.u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
// (sample1 + sample2 + offset)>>shift
temp_u_epi16 = _mm256_add_epi16(sample0_epi16, sample1_epi16);
@ -1120,10 +1064,10 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
temp_u_epi16 = _mm256_srai_epi16(temp_u_epi16, shift);
sample0_epi16 = hi_prec_chroma_rec0 ? _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*) &(high_precision_rec0->v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))) :
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_loadl_epi64((__m128i*) &(temp_lcu_v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_cvtsi32_si128(*(int32_t*) &(temp_lcu_v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
sample1_epi16 = hi_prec_chroma_rec1 ? _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*) &(high_precision_rec1->v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))) :
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_loadl_epi64((__m128i*) &(lcu->rec.v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
_mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm_cvtsi32_si128(*(int32_t*) &(lcu->rec.v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]))), shift_left);
// (sample1 + sample2 + offset)>>shift
temp_v_epi16 = _mm256_add_epi16(sample0_epi16, sample1_epi16);

16
tests/inter_recon_bipred_tests.c
View file

@ -32,9 +32,9 @@ static lcu_t lcu1;
int temp1, temp2, temp3, temp4;
int16_t mv_param[2][2] = { { 3,3 },{ 3,3 } };
int width = 32;
int height = 32;
int16_t mv_param[2][2] = { { 4,4 },{ 4,4 } };
int width = 4;
int height = 4;
int xpos = 0;
int ypos = 0;
@ -121,11 +121,11 @@ static void setup()
TEST test_inter_recon_bipred()
{
/*
memcpy(result.rec.y, lcu1.rec.y, sizeof(kvz_pixel) * 64 * 64);
memcpy(result.rec.u, lcu1.rec.u, sizeof(kvz_pixel) * 32 * 32);
memcpy(result.rec.v, lcu1.rec.v, sizeof(kvz_pixel) * 32 * 32);
/*
for (temp_y = 0; temp_y < height; ++temp_y) {
int y_in_lcu = ((ypos + temp_y) & ((LCU_WIDTH)-1));
for (temp_x = 0; temp_x < width; temp_x += 1) {
@ -133,7 +133,7 @@ TEST test_inter_recon_bipred()
printf("%d ", expected_test_result.rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu]);
}
}
printf("\n");*/
printf("\n");
/*
for (temp_y = 0; temp_y < height >> 1; ++temp_y) {
int y_in_lcu = (((ypos >> 1) + temp_y) & (LCU_WIDTH_C - 1));
@ -144,7 +144,7 @@ TEST test_inter_recon_bipred()
}
printf("\n");*/
kvz_inter_recon_bipred_generic(hi_prec_luma_rec0, hi_prec_luma_rec1, hi_prec_chroma_rec0, hi_prec_chroma_rec1, width, height, xpos, ypos, high_precision_rec0, high_precision_rec1, &result, temp_lcu_y, temp_lcu_u, temp_lcu_v);
kvz_inter_recon_bipred_blend(hi_prec_luma_rec0, hi_prec_luma_rec1, hi_prec_chroma_rec0, hi_prec_chroma_rec1, width, height, xpos, ypos, high_precision_rec0, high_precision_rec1, &result, temp_lcu_y, temp_lcu_u, temp_lcu_v);
/*
for (temp_y = 0; temp_y < height; ++temp_y) {
int y_in_lcu = ((ypos + temp_y) & ((LCU_WIDTH)-1));
@ -195,7 +195,7 @@ SUITE(inter_recon_bipred_tests)
continue;
}
kvz_inter_recon_bipred_generic = strategies.strategies[i].fptr;
kvz_inter_recon_bipred_blend = strategies.strategies[i].fptr;
RUN_TEST(test_inter_recon_bipred);
}
}